知识卡片
数据歧义不构成取消验证的理由
训练57%、验证29%可能涉及重叠标签与分布问题,但仍应保留独立评价并修正目标、分组和指标。与25类均匀随机的4%比较只是一种基准,还需多数类与类别分布;不能据此证明达到人类专家水平。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- · 查看原书依据
It looks like the RNN achieved 57% accuracy on the training set and 29% accuracy on the validation set. This is an unfair measure of the model’s usefulness. Because the dataset was deduplicated before splitting into training and validation sets, there is only one row in the dataset for each name–nationality combination. This means a name associated with one nationality in the training set will
参考来源
- 位置:《Natural Language Processing in Action, Second Edition》第08章《循环网络与长短期记忆》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-8.html)
- 结论依据:摘录中关于数据歧义不构成取消验证的理由的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:It looks like the RNN achieved 57% accuracy on the training set and 29% accuracy on the validation set. This is an unfair measure of the model’s usefulness. Because the dataset was deduplicated before splitting into training and validation sets, there is only one row in the dataset for each name–nationality combination. This means a name associated with one nationality in the training set will